Can you tell a fake face from a real one?
Researchers at University of Washington created a website, whichfaceisreal.com, to quiz people on whether they can tell a real person’s face from a computer-generated one.
Jevin West and Carl Bergstrom, the site’s creators, told BuzzFeed News the goal is to help people understand how to spot fakes created by computers. They also teach a course, "Calling Bullshit," that helps students “think critically about the data and models that constitute evidence in the social and natural sciences.”
“We’re trying to teach them how to spot BS in a graphic or spot BS in an image,” West said.
The website is a passion project, but it highlights real concerns with machine-generated images and videos. Deepfakes, a term used to describe artificially generated videos of real people, could be used to spread disinformation and are already being deployed to harass women.
“Governments used to worry about counterfeiting money; now we have to worry about counterfeiting people,” Bergstrom said.
Throwaway email addresses could be combined with fake images of people, for example, to fabricate entire online personas. Bergstrom cited examples like the fake FTC comments submitted to the net neutrality consultation process, and Facebook having to remove 1.3 billion fake profiles from its website.
“It’s an enormous threat to democracy if you can counterfeit opinions of people,” he said.
There are a handful of telltale signs you’re looking at a computer-generated face, which West and Bergstrom lay out on their website. The key is to look for imperfections like water splotches and asymmetries. Glasses, teeth, and backgrounds are all more difficult to render. So are earrings, especially intricate ones. If the person in the image is wearing intricate, identical earrings it’s likely that person is real.
West and Bergstrom said after people take the quiz a few times they get better at telling the difference between real and fake faces. But that skill could become obsolete in three years or so, when machine learning advances enough to iron out the imperfections. Even now, computer-generated people pose a problem.
“The combination of AI and someone with some Photoshop skill would make it really hard to detect,” Bergstrom said. “You could clean up most of the problems. Luckily, right now we can’t do it to scale.”